A scalable geometric algorithm for community detection from social networks with incremental update

Abstract

In recent years, a series of algorithms have been proposed to detect community from social networks. Most of the algorithms are based on traditional spectral clustering algorithms such as k-means. One of the major limitations of such algorithms is that entire eigenvalues of the similarity matrix of the network need to be calculated in advance. In the case of a massive network, calculating entire eigenvalues is computationally expensive. This paper proposes a scalable geometric algorithm to find communities from large social networks. The major contributions of this work are: (1) We transform the network data into points by preserving the intrinsic properties and structure of the original data. (2) A novel geometric clustering is derived. And we use the data structure C-Tree and Voronoi diagram for identifying communities from the points in the Euclidean plane. (3) Since social networks grow dynamically, we further extend the algorithm to incrementally identify the community membership of newly introduced members. Experiments on both synthetic and real-world datasets show that the algorithm, in terms of objective matrices, is equally good as spectral clustering algorithm.